Analysis: AI Agents in Banking Hit a Liability Wall

Citi, PNC and Citizens told customers this week that they are responsible for whatever their AI agents do in their accounts, and Citi even covers leaked credentials. Singapore's central bank wants a know your agent protocol instead. We explain why banks are closing the side door, who wins and loses, and what to watch as the first sanctioned agent channels arrive.

Christopher Ort
By i10X News Desk · Reviewed by Christopher Ort · 12 min read
Share
A heavy steel bank vault door standing slightly ajar in a dim marble hall, with a small glowing robotic hand-shaped silhouette of light reaching toward the gap and blurred blue screen light behind.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened this week
  4. Why it matters
  5. Context: how we got here
  6. The data: who says what
  7. Winners and losers
  8. What people are saying on X
  9. Three scenarios for the next year
  10. What to watch next
  11. About the analysis
  12. i10x Perspective

In the space of one week, the biggest open question about AI agents and money got a set of answers, and almost all of them point the same way. Citi rewrote its Online User Agreement so that anything a third-party AI agent does with a customer’s login counts as the customer’s own act, even after the credentials leak. PNC published new terms that forbid giving an agent your password at all. Singapore’s central bank said it will build a “know your agent” protocol and told consumers not to hand personal agents their bank logins. And in Arizona, the attorney general opened a criminal investigation into what she called rogue AI agents inside banks. Together these moves define how AI agents in banking will work for the next few years: banks are closing the side door and starting to build a front door, and until it opens, the user carries the risk.

i10X Quick Take

Quick Take

  1. What happened: Citi (Section H of its Online User Agreement), PNC (a new Digital Services Agreement effective December 19, 2026) and Citizens (a customer email reported by TechCrunch) all told customers this week that they are responsible for what their AI agents do in their accounts.
  2. The sharpest clause: Citi treats actions taken with an agent’s credentials as authorized by the customer even if those credentials are “further shared, leaked, or otherwise distributed,” and says it has no duty to investigate.
  3. The front door: Singapore’s MAS will facilitate a know your agent protocol so banks can verify who an agent is, whom it represents, what authority it has and under what mandate it acts. JPMorgan’s Kinexys tested exactly that model with Hyundai Card.
  4. The pressure: Arizona’s criminal probe and a Federal Reserve governor’s warning that agents are already acting in retail accounts without verified consent push banks toward stricter control.
  5. Our read: Credential sharing with agents is being shut down by contract. The winners will be whoever owns the verified, permissioned channel that replaces it.

In plain English

Many AI assistants can now log into websites for you. Some people have started giving them their online banking username and password so the assistant can check balances, pay bills or move savings. Banks cannot tell whether the person logging in is you or software acting for you, or whether that software is doing what you actually asked. So the big banks are changing their rules: if you give an AI tool your login, whatever it does is on you, including mistakes and what happens if the tool is hacked. At the same time, regulators and banks are working on a safer system in which an agent gets its own ID and a limited permission slip instead of your password.

What happened this week

On November 18, TechCrunch reported that Citi had added a new section to its Online User Agreement covering third-party AI agents. We covered the text in our report on Citi’s new AI agent terms. Section H defines an AI agent as “software or other technology that takes autonomous, semi-autonomous, or non-autonomous action on behalf of, or at the instruction of, You.” It cautions customers against using any third-party agent and makes them responsible for “any actions or errors that result from the third-party AI Agent’s incorrect interpretation or execution of your instructions, unintended actions performed by the third-party AI Agent, or actions taken by the third-party AI Agent that exceed the authority granted.” The terms do not apply to AI tools Citi provides itself.

A Citi spokesperson told TechCrunch: “As third-party AI tools become more widely available, we want customers to understand the risks associated with sharing banking credentials or authorizing those tools to access their accounts.” The bank said the update reinforces “that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents.”

PNC went a step further on the technical side. Its updated Digital Services Agreement, effective December 19, 2026, has a new section titled “Third-Party AI Agents.” It says customers “must not disclose, share, enter, or otherwise provide” their User ID, password or one-time passcodes to any third-party AI agent. Agents may only connect through methods PNC makes available, and unless PNC expressly permits it for a particular service, a customer “may not authorize or permit a Third-Party AI Agent to initiate, request, submit, or complete a payment, transfer, or other transaction.” The agreement also bans letting agents use “browser automation, screen scraping, an embedded or shared browser session, or similar technology” to get around those restrictions. According to TechCrunch, Citizens emailed customers this week with a similar warning.

On November 19, Monetary Authority of Singapore Managing Director Chia Der Jiun used his keynote at the Singapore FinTech Festival to describe the problem from the regulator’s side, as we reported in our coverage of the MAS know your agent plan. “Some users are giving agents their login credentials to financial accounts. I would not advise giving personal agents such access, as there is insufficient assurance of safety,” he said. He named three gaps: the institution does not know whether the agent acts with the customer’s full authorization, whether it stays within its mandate, and “if agents transact incorrectly or beyond their instruction, who bears responsibility?”

Why it matters

For two years, the pitch for consumer AI agents has included money: an assistant that finds a better savings rate, cancels unused subscriptions and pays bills on time. Technically, the shortcut was simple. The agent logs in with the user’s own credentials and clicks through the bank’s website like a person would. This week’s terms make that shortcut legally expensive for the user and, in PNC’s case, a breach of contract.

That matters because it shifts the question from “can an agent do this?” to “who is liable when it does?” Under Citi’s terms, a customer whose agent misreads an instruction, overspends or gets compromised has accepted the outcome in advance, subject to whatever consumer protection law still applies. Citi’s agreement keeps that caveat, stating customer responsibility applies “except as otherwise permitted under applicable law.” How those clauses interact with existing US rules on unauthorized electronic transfers has not been tested in court.

It also matters for AI companies. A consumer agent that cannot touch a bank account is far less useful as a personal finance tool. Anthropic’s Personal Finance in Claude, launched on November 5, was designed for this world from the start: it links accounts through Plaid, is read-only and, in Anthropic’s words, cannot move, manage or invest money. OpenAI’s ChatGPT Wallet, introduced on November 13, takes the other sanctioned route for payments: the agent never sees the card number, the card network issues a payment token, and the user approves each checkout with a biometric check.

Context: how we got here

The crackdown did not start this week. Between October 13 and October 20, according to a legal update by the law firm Hunton Andrews Kurth, JPMorgan Chase, Charles Schwab, Citigroup and Wells Fargo each moved to block or contractually reallocate the risk of third-party agents logging in with customer credentials. JPMorgan CEO Jamie Dimon said on the bank’s third-quarter earnings call that it would not allow outside agents to act inside customer accounts, citing “rogue agents.” Schwab CEO Rick Wurster told Bloomberg that the firm blocks unauthorized agents that try to move client money 10 to 20 times per week. We described that first wave in our report on the AI agent identity crisis at banks.

What changed this week is the step from blocking to writing liability into contracts, and from banks acting alone to a central bank proposing shared infrastructure. Regulators had been signaling concern. On November 16, Federal Reserve Governor Michael Barr said that “major banks now report finding AI agents attempting to take actions in retail customer accounts, without the bank being able to verify authenticity, consent, and alignment with customer instructions.” He called for “strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities,” and warned that agents moving funds at high speed could accelerate bank runs.

The legal environment also tightened. On November 18, Arizona Attorney General Kris Mayes announced a criminal investigation into OpenAI and other AI companies over the “apparent infiltration of banks by rogue AI agents,” Reuters reported. No charges have been filed, the banks involved were not named, and an investigation does not necessarily lead to charges. Two days earlier, California’s attorney general had sued OpenAI over rogue agent incidents.

There is a familiar pattern underneath. The same banks spent years fighting credential-based screen scraping by data aggregators, which led to token-based, permissioned connections for apps such as budgeting tools. The rule meant to standardize that access in the US, the Consumer Financial Protection Bureau’s Section 1033 open banking rule, is still unsettled, according to Hunton. Every bank that has acted so far draws the same line: read-only data through approved connectors is allowed, credential logins and money movement by outside agents are not.

The data: who says what

InstitutionDocument or statementWhat it says about AI agentsDate
CitiOnline User Agreement, Section HAgent actions count as authorized by the customer, including after credentials leak; no duty to investigate; Citi’s own AI tools excludedUpdated November 2026
PNCDigital Services Agreement, Third-Party AI AgentsNo sharing of logins or passcodes with agents; agents only via PNC connection methods; no payments or transfers unless PNC expressly permits; no browser automation or screen scrapingEffective December 19, 2026
CitizensCustomer email (reported by TechCrunch)Customers are responsible for what their AI assistant doesNovember 2026
Wells FargoOnline Access Agreement, Section 15 (per Hunton)Use at your own risk; agent-initiated transactions deemed authorized; no duty to monitorEffective November 24, 2026
JPMorgan ChaseDigital Services Agreement update (per Hunton)A customer who gives an agent access authorizes everything it doesEffective January 1, 2027
MAS (Singapore)Keynote at Singapore FinTech FestivalKnow your agent protocol covering identity, authority, mandate and assurance; advises users not to share loginsAnnounced November 19, 2026
Kinexys by J.P. Morgan and Hyundai CardProof of conceptBank-side agent checks a cryptographic agent identity and client-approved mandate before acting; simulated money onlyAnnounced November 18, 2026

The table shows a clear split. US banks are defining agents as a customer risk and pushing liability outward. MAS and the Kinexys pilot are defining agents as a new kind of authorized party that needs identity, a mandate and controls. Both approaches end in the same place: an agent that wants to move money will have to arrive through a verified channel, not with a borrowed password.

Banks are not anti-AI in their own operations. Our Banks Using AI tracker lists 30 deployments at 28 banks and fintechs in 13 countries, most of them employee assistants, including Citi Assist and Citi Stylus for more than 175,000 Citi employees and JPMorgan’s LLM Suite for more than 200,000. The banks tightening the rules for outside agents are among the heaviest internal users of AI. Citi’s new terms explicitly exempt the AI tools it provides itself.

Winners and losers

Winners. Banks gain legal cover and a stronger hand in deciding which agents reach their customers, which TechCrunch notes is also a competitive question, since agents that move deposits to higher-yielding accounts threaten bank margins. Providers of permissioned connections and agent identity, such as Plaid-style aggregators, token-based card networks and bank-side infrastructure like Kinexys, become the toll roads for agentic finance. Agent builders that already designed for read-only data or tokenized payments, such as Anthropic’s finance tool and ChatGPT Wallet, face little disruption.

Losers. General-purpose browser agents that rely on logging in as the user lose their most valuable consumer use case in US banking. Users who already connected agents with their passwords carry new risk, often without having read the updated terms. Fintech startups building autonomous money managers on top of screen scraping now face both technical blocks and contractual bans.

Still undecided. Regulators. MAS has chosen to build shared standards with industry rather than prescriptive rules. In the US, the authorization and liability questions raised by agent-initiated transactions remain open, and Arizona’s probe adds a criminal law dimension that no bank or AI company controls.

What people are saying on X

Coverage of the MAS announcement on X focused on the know your agent protocol as part of Singapore’s wider push on digital settlement infrastructure.

Investors picked up on the Citi change as part of a broader trend rather than a one-off.

Three scenarios for the next year

1. The front door opens (most likely). Banks keep blocking credential logins but launch scoped agent connections, starting with read-only data and moving to limited payments with per-transaction approval. Standards such as the MAS know your agent protocol and card network payment tokens converge into something banks can plug into. Liability clauses stay in place but matter less, because sanctioned channels come with their own rules.

2. A long standoff. Banks keep the contracts and blocks, agent builders keep finding workarounds, and users keep sharing passwords. The first large loss caused by an agent with leaked credentials becomes a test case for how far contractual liability can go against consumer protection law.

3. Regulators set the terms. A US open banking rule, supervisory guidance or the outcome of state actions such as Arizona’s investigation forces a common standard on agent authorization and liability, overriding individual bank terms. This would be slower but would give users the clearest protection.

What to watch next

  • Effective dates. Wells Fargo’s new AI agent section takes effect on November 24, 2026, PNC’s on December 19, 2026, and JPMorgan’s updated Digital Services Agreement on January 1, 2027, according to Hunton and PNC’s own document.
  • The first sanctioned agent channels. PNC’s terms already describe a connection method where customers authenticate directly with the bank and set accounts, duration and permissions. Watch which banks turn that into a product and whether payments are included.
  • MAS partners and timelines. Which banks and technology companies join the know your agent work and the Agentic Money Sandbox with the Future of Finance Institute.
  • Arizona’s investigation. Which companies and banks are involved and which statutes are in question.
  • Agent identity proposals from AI companies. Hunton notes that Meta is developing a protocol called Poppy to signal when a personal AI agent is acting for a customer. Whether banks accept such signals will decide how open the front door is.

About the analysis

This analysis is based on Citi’s Online User Agreement, Section H “Use of Third-party Artificial Intelligence (‘AI’) Agents” (updated November 2026); PNC’s Digital Services Agreement, version effective December 19, 2026, section “Third-Party AI Agents”; the keynote address by MAS Managing Director Chia Der Jiun at the Singapore FinTech Festival 2026 on November 19, 2026; TechCrunch’s report by Tim Fernholz of November 18, 2026, including Citi’s spokesperson statement and the Citizens email; the Hunton Andrews Kurth legal update “Major Banks Move to Block or Limit Customer-Directed AI Agents” of November 3, 2026; Governor Michael Barr’s speech of November 16, 2026; Reuters’ report on the Arizona investigation of November 18, 2026; Anthropic’s and OpenAI’s product announcements; and posts on X by CoinDesk and M Mohan. Deployment figures come from the i10X Banks Using AI tracker. This article is analysis, not legal or financial advice.

i10x Perspective

AI agents were supposed to make money management effortless. This week showed that the bottleneck is not intelligence but trust plumbing: identity, mandates, audit trails and a clear answer to who pays when something goes wrong. For now, banks have answered that last question in their own favor, and users should take it literally: do not give an AI agent your banking password. The more interesting race is the one that follows. Whoever defines how an agent proves who it is and what it may do, whether a central bank protocol, a card network token or a bank’s own connector, will shape how agents touch money for the next decade. We will keep tracking every bank that opens, or closes, that door.

Christopher Ort

Christopher Ort

Christopher Ort is the editor of i10X News at i10x.ai. He has worked hands-on with AI tools, models and agent workflows for years and leads i10X’s coverage of AI launches, funding and research. He is the author of “The Style Penalty”, an SSRN study on how frontier AI models judge AI-written CVs, covered by heise online and t3n.

Editor's picks

  1. Dragos canary study: attackers stole AI prompts and credentials at three AI inference providers
  2. Analysis: AI Agents in Banking Hit a Liability Wall
  3. CVE-2026-133631: FlowIntel chatbot let LLM call any API, exposed Ollama token in process table
  4. Stripe Evals: Coding Agents Score 78% on Payment Integrations, but Only 29% Pass Every Required Check
  5. Kong Volcano: A Self-Healing Agent Cloud Where Coding Agents Build, Host and Fix Apps